Exploring the identification of multiple bacteria on stainless steel using multi-scale spectral imaging from microscopic to macroscopic.

This work investigates non-contact reflectance spectral imaging techniques, i.e. microscopic Fourier transform infrared (FTIR) imaging, macroscopic visible-near infrared (VNIR), and shortwave infrared (SWIR) spectral imaging, for the identification of bacteria on stainless steel. Spectral images of two Gram-positive (GP) bacteria (Bacillus subtilis (BS) and Lactobacillus plantarum (LP)), and three Gram-negative (GN) bacteria (Escherichia coli (EC), Cronobacter sakazakii (CS), and Pseudomonas fluorescens (PF)), were collected from dried suspensions of bacterial cells dropped onto stainless steel surfaces. Through the use of multiple independent biological replicates for model validation and testing, FTIR reflectance spectral imaging was found to provide excellent GP/GN classification accuracy (> 96%), while the fused VNIR-SWIR data yielded classification accuracy exceeding 80% when applied to the independent test sets. However, classification within gram type was far less reliable, with lower accuracies for classification within the GP (< 75%) and GN (≤ 51%) species when calibration models were applied to the independent test sets, underlining the importance of independent model validation when dealing with samples of high biological variability.

Detection of previously frozen poultry through plastic lidding film using portable visible spectral imaging (443-726 NM).

The objective of this study is to use a portable visible spectral imaging system (443-726 nm) to detect poultry thawed from frozen at the pixel level using multivariate analysis methods commonly used in machine learning (decision tree, logistic regression, linear discriminant analysis [LDA], k-nearest neighbors [KNN], support vector machines [SVM]). The selection of the most suitable method is based on the amount of data required to build an accurate model, computational speed, and the robustness of the model. The training set consists of pixel spectra from packages of chicken thighs without plastic lidding to evaluate the robustness of the models when implemented on the test set with and without plastic lidding. Data subsets were created by randomly selecting 1, 5, 10, 20, and 50% of the pixel spectra of each sample for both the training and test data sets. The subsets of pixel spectra and the full training set were used to train the machine learning algorithms to evaluate how the amount of data influences computational time. Logistic regression was found to be the best algorithm for detecting poultry thawed from frozen with and without plastic lidding film. Although logistic regression and SVM both performed with the same high accuracy and sensitivity for all training subset sizes, the computational time needed to implement SVM makes it the less suitable algorithm for detecting poultry thawed from frozen with and without plastic lidding film.